Comprehensive engineering course covering the end-to-end lifecycle of production deep learning and LLM systems: data management, experiment tracking, troubleshooting, testing, deployment, and continual learning.
Taught by Sergey Karayev, Josh Tobin, and Pieter Abbeel (originating at UC Berkeley), Full Stack Deep Learning focuses on the critical engineering practices that turn research models into reliable systems: setting up reproducible project structures, managing datasets, diagnosing training failures, running systematic ablations, CI/CD testing for ML models, and deploying/monitoring models in production.
Course status is tracked independently from roadmap topic checkboxes so you can mark external courses without altering your topic mastery percentage.
Engineering Lifecycle Course across Phase 6 (Training & Packaging) and Phase 8 (Deployment & Systems)
All lecture videos, full written lecture notes, slides, and Jupyter lab repositories are 100% free and openly accessible.
Status: Verified (2026-11-09)
Review these foundations before starting so you get the most out of this material:
In Phase 6: Train Your Own LLM
Use the lectures on Data Management, Troubleshooting & Testing, and Experiment Management during Modules 6.3–6.5 to structure your `MyLLM v1` training runs cleanly.
In Phase 8: Advanced LLM Systems & Research
In Phase 8, study the lectures on Infrastructure & Tooling, Deployment, Continual Learning, and Foundation Model Systems alongside Modules 8.4 and 8.5.
Corpus curation, licensing, cleaning, deduplication, PII filtering, tokenization, sequence packing, and memory-mapped binary streaming.
Progressive pretraining experiments across Model A (100K–1M params), Model B (5M–20M params), and Model C (30M–100M params), tailored to your available hardware.
Training vs. validation loss, perplexity, baselines, memorization vs. generalization, data contamination, hallucinations, robustness, speed/memory profiling, and ablations.
Serializing model weights, tokenizer artifacts, configuration JSONs, model cards, and standalone CLI inference scripts for reproducible distribution.
Post-training quantization (INT8, INT4, AWQ, GGUF), KV cache management, continuous batching, PagedAttention, speculative decoding, serving APIs, observability, and cost optimization.
Reading AI research papers critically, formulating testable hypotheses, establishing fair baselines, controlled ablations, statistical rigor, and reproducing published results.
Graduate-level Stanford course walking through every aspect of building language models from scratch: Transformer architecture, tokenization, pretraining data pipelines, GPU systems, parallelism, and scaling laws.
Official documentation for vLLM, the high-throughput open-source LLM inference and serving engine powered by PagedAttention, continuous batching, speculative decoding, and quantized kernels.
Official PyTorch documentation and step-by-step tutorials for DistributedDataParallel (DDP), Fully Sharded Data Parallel (FSDP2), Tensor Parallelism, and multi-GPU collective communication.